one small box, measured in public

a base M4 mac mini with 24GB runs my local models and agents all day. real tok/s, real costs, every flag.

hailports is one standard M4 mac mini with 24GB of unified memory. no pro chip, no egpu. it runs a 35B mixture-of-experts model locally through llama.cpp, plus the agents that use it all day for real work.

what the agents do

they identify the task and carry out the work. a loop watches queues, inboxes and logs, spots what needs doing, drafts or does it, checks it against gates, and ships it into a review queue. a human approves and decides. anything that spends money, sends something or changes strategy waits for a yes.

• 16 days, deduped: 2,075 unique work items, about 2,633 local model calls a day
• paid frontier API burn in the Nov 3, 2026 write-up: $42.50 a day before, $3.90 a day after (-91%)
• Nov 6, 2026, 11:58 AM CT: daily median 43.5 tok/s across 1,217 answers of 100+ tokens; fastest 54.9 tok/s. dated measurements on the live board

the rules this site follows

• every number comes from the box's logs, with its date next to it. no invented figures, no testimonials, no user counts
• speed claims carry their measurement date and sample size. the Nov 2 write-up is historical; the live board has the latest measured median
• the box is a base M4 with 24GB. it has never been a pro
• the operator stays anonymous. the work is the proof

where to start

the 41 tok/s write-up has the exact launch line. the cost write-up prices 16 days of local traffic at api rates. the live board has today's numbers.